Adaptive-Preprocessing-Enhanced YOLOX-L For Breast Cancer Mass Detection in Contrast-Enhanced Spectral Mammography

Abstract

Breast cancer detection using contrast-enhanced spectral mammography (CESM) is still a challenging task, especially when lesions are subtle, low in contrast, or located within dense breast tissue. In this paper, an automated breast mass localization pipeline was developed using the YOLOX-L object detection framework. The proposed pipeline applies a piecewise adaptive contrast enhancement method before the detection stage. This method depends on dynamic thresholding to improve the visibility of lesion regions while preserving the surrounding anatomical structures. The proposed method was evaluated using the public CDD-CESM dataset. The enhanced YOLOX-L model achieved a mAP@0.50:0.95 of 0.703, while the unenhanced baseline achieved 0.186 under the same evaluation setting. Furthermore, additional experiments showed that the proposed enhancement method performed better than histogram equalization and CLAHE when using the same patient-level split, training configuration, and evaluation protocol. The pipeline also achieved near-real-time inference under the tested computational environment. Hence, the proposed system may be useful as an assistive tool for CESM breast mass detection. However, further external and prospective validation is still required before clinical use.